Apache Spark 2.4 comes packed with a lot of new functionalities and improvements, including the new barrier execution mode, flexible streaming sink, the native AVRO data source, PySpark’s eager evaluation mode, Kubernetes support, higher-order functions, Scala 2.12 support, and more.

1. Apache Spark 2.4 and Beyond
Xiao Li, Wenchen Fan
Mar 2019 @ Strata Data Conf

2. About US
• Software Engineers at
• Apache Spark Committers and PMC Members
Xiao Li (Github: gatorsmile) Wenchen Fan (Github: cloud-fan)

3. Databricks Customers Across Industries
Financial Services Healthcare & Pharma Media & Entertainment Technology
Public Sector Retail & CPG Consumer Services Energy & Industrial IoTMarketing & AdTech
Data & Analytics Services

4. DATABRICKS WORKSPACE
Databricks Delta ML Frameworks
DATABRICKS CLOUD SERVICE
DATABRICKS RUNTIME
Reliable & Scalable Simple & Integrated
Databricks Unified Analytics Platform
APIs
Jobs
Models
Notebooks
Dashboards End to end ML lifecycle

6. https://insights.stackoverflow.com/survey/2018

7. 20000+ Stars in Github

8. Top Skill in 2018: Apache Spark
https://economicgraph.linkedin.com/research/linkedin-2018-emerging-jobs-report

10. 10
Release: Nov 8, 2018
Blog: https://t.co/k7kEHrNZXp
Above 1100 tickets.

11. Higher-order
Functions
Major Features on Spark 2.4
11
Structured
Streaming
Built-in source
Improvement
Spark on
Kubernetes
PySpark
Improvement
Native Avro
Support
Image
Source
Barrier
Execution
Beta support
Scala 2.12
Various SQL
Features

12. Survey Key Findings: The AI Dilemma
Investment in AI is Growing Quickly
However, very few are succeeding
Major contributing factors to this AI dilemma
7
Different machine learning
frameworks and tools
https://databricks.com/CIO-survey-report

13. CIO Survey Key Findings: Unified Analytics Enables AI Success
Unifying data science & engineering
Benefits Expected from a Unified Approach to Data & AI
Increased operational efficiency
More effective decision making
Accelerated time to market
Improved security
Increased innovation
Improved customer experience
Increased competitive advantage
Increased employee satisfaction
Increased customer engagement
Product/service transformation
Topline growth
Other (specify)

14. Big data v.s. AI Technologies
14
X

15. Project Hydrogen: Spark + AI
A gang scheduling to Apache Spark that embeds a
distributed DL job as a Spark stage to simplify the
distributed training workflow. [SPARK-24374]
15
Task 1
Task 2
Task 3

16. Higher-order
Functions
Major Features on Spark 2.4
16
Structured
Streaming
Built-in source
Improvement
Spark on
Kubernetes
PySpark
Improvement
Native Avro
Support
Image
Source
Barrier
Execution
Scala
2.12
Various SQL
Features

17. Flexible Streaming Sink
[SPARK-24565] Exposing output rows of each microbatch as a
DataFrame
foreachBatch(f: Dataset[T] => Unit)
• Scala/Java/Python APIs in DataStreamWriter.
• Reuse existing batch data sources
• Write to multiple locations
• Apply additional DataFrame operations
17

18. Reuse existing batch data sources
18

19. Write to multiple location
19

20. Higher-order
Functions
Major Features on Spark 2.4
20
Structured
Streaming
Built-in source
Improvement
Spark on
Kubernetes
PySpark
Improvement
Native Avro
Support
Image
Source
Barrier
Execution
Scala
2.12
Various SQL
Features

21. Parquet
21
Update from 1.8.2 to 1.10.0 [SPARK-23972].
• PARQUET-1025 - Support new min-max statistics in parquet-mr
• PARQUET-225 - INT64 support for delta encoding
• PARQUET-1142 Enable parquet.filter.dictionary.enabled by default.
Predicate pushdown
• STRING [SPARK-23972] [20x faster]
• Decimal [SPARK-24549]
• Timestamp [SPARK-24718]
• Date [SPARK-23727]
• Byte/Short [SPARK-24706]
• StringStartsWith [SPARK-24638]
• IN [SPARK-17091]

22. ORC
Native vectorized ORC reader is GAed!
• Native ORC reader is on by default [SPARK-23456]
• Update ORC from 1.4.1 to 1.5.2 [SPARK-24576]
• Turn on ORC filter push-down by default [SPARK-21783]
• Use native ORC reader to read Hive serde tables by default
[SPARK-22279]
• Avoid creating reader for all ORC files [SPARK-25126]
22

23. Higher-order
Functions
Major Features on Upcoming Spark 2.4
23
Structured
Streaming
Built-in source
Improvement
Spark on
Kubernetes
PySpark
Improvement
Native Avro
Support
Image
Source
Barrier
Execution
Scala
2.12
Various SQL
Features

24. Higher-order Functions
Transformation on complex objects like arrays, maps and structures
inside of columns.
24
tbl_nested
|-- key: long (nullable = false)
|-- values: array (nullable = false)
| |-- element: long (containsNull = false)
UDF ? Expensive data serialization

25. 1) Check for element existence
SELECT EXISTS(values, e -> e > 30) AS v
FROM tbl_nested;
2) Transform an array
SELECT TRANSFORM(values, e -> e * e) AS v
FROM tbl_nested;
tbl_nested
|-- key: long (nullable = false)
|-- values: array (nullable = false)
| |-- element: long (containsNull = false)
Higher-order Functions

26. 4) Aggregate an array
SELECT REDUCE(values, 0, (value, acc) -> value + acc) AS sum
FROM tbl_nested;
Ref Databricks Blog: http://dbricks.co/2rUKQ1A
3) Filter an array
SELECT FILTER(values, e -> e > 30) AS v
FROM tbl_nested;
tbl_nested
|-- key: long (nullable = false)
|-- values: array (nullable = false)
| |-- element: long (containsNull = false)
Higher-order Functions

27. Built-in Functions
[SPARK-23899] New or extended built-in functions for ArrayTypes and
MapTypes
• 26 functions for ArrayTypes
transform, filter, reduce, array_distinct, array_intersect, array_union,
array_except, array_join, array_max, array_min, ...
• 3 functions for MapTypes
map_from_arrays, map_from_entries, map_concat
27
Blog: Introducing New Built-in and Higher-Order Functions for
Complex Data Types in Apache Spark 2.4. https://t.co/p1TRRtabJJ

28. Higher-order
Functions
Major Features on Spark 2.4
28
Structured
Streaming
Built-in source
Improvement
Spark on
Kubernetes
PySpark
Improvement
Native Avro
Support
Image
Source
Barrier
Execution
Scala
2.12
Various SQL
Features

29. Native Spark App in K8S
New Spark scheduler backend
• PySpark support [SPARK-23984]
• SparkR support [SPARK-24433]
• Client-mode support [SPARK-23146]
• Support for mounting K8S volumes
[SPARK-23529]
Blog: What’s New for Apache Spark on Kubernetes in
the Upcoming Apache Spark 2.4 Release
https://t.co/uUpdUj2Z4B
29
on

30. Higher-order
Functions
Major Features on Spark 2.4
30
Structured
Streaming
Built-in source
Improvement
Spark on
Kubernetes
PySpark
Improvement
Native Avro
Support
Image
Source
Barrier
Execution
Beta support
Scala 2.12
Various SQL
Features

31. What’s Next?

32. This presentation may contain projections or other forward-
looking statements regarding the upcoming release (Apache
Spark 3.0). The statements are intended to outline our general
direction. They are intended for information purposes only.
They are not a commitment to deliver code or functionality.
The development, release and timing of any feature or
functionality described for Apache Spark remains at the sole
discretion of ASF and the Apache Spark PMC.
Safe Harbor Statement

33. What’s Next?
33
Data Source
APIs
Spark on
Kubernetes
PySpark
Usability
Scala 2.12 Various SQL
Features
GPU-aware
Scheduling
Adaptive
Execution
Hadoop 3.x
Spark Graph
mlflow

34. What’s Next?
34
Data Source
APIs
Spark on
Kubernetes
PySpark
Usability
Scala 2.12 Various SQL
Features
GPU-aware
Scheduling
Adaptive
Execution
Hadoop 3.x
Spark Graph
mlflow

35. Project Hydrogen: Spark + AI
GPU Aware Scheduling
• widely used for accelerating special workloads, e.g., deep
learning and signal processing
35

36. It’s Hard to Productionize ML

37. ML Lifecycle is Manual, Inconsistent
and Disconnected
● Ad hoc approach to track
experiments
● Very hard to reproduce
experiments
Data Prep
● Multiple tightly coupled
deployment options
● Different monitoring approach
for each framework
Build Model Deploy Model
● Low level integrations for
Data and ML
● Difficult to track data used
for a model

38. What is ?
Open source platform to manage ML development
• Lightweight APIs & abstractions that work with any ML library
• Designed to be useful for 1 user or 1000+ person orgs
• Runs the same way anywhere (e.g. any cloud)
Key principle: “open interface” APIs that work with any existing
ML library, app, deployment tool, etc

39. MLflow Components
39
Tracking
Record and query
experiments: code,
params, results, etc
Projects
Code packaging for
reproducible runs
on any platform
Models
Model packaging and
deployment to diverse
environments

40. MLflow Projects
• Docker-based project environment
specification (0.9)
• X-coordinate logging for metrics &
batched logging (1.0)
• Packaging projects with build steps (1.0+)
MLflow Models
• Custom model logging in Python, R
and Java (0.8, 0.9, 1.0)
• Better environment isolation when
loading models (1.0)
• Logging schema of models (1.0+)
MLflow Tracking
• SQL database backend for scaling the tracking server (0.9)
• UI scalability improvements (0.8, 0.9, etc.)
• X-coordinate logging for metrics & batched logging (1.0)
• Fluent API for Java and Scala (1.0)
What’s
Next?

41. 41

42. What’s Next?
42
Data Source
APIs
Spark on
Kubernetes
PySpark
Usability
Scala 2.12 Various SQL
Features
GPU-aware
Scheduling
Adaptive
Execution
Hadoop 3.x
Spark Graph
mlflow

43. Challenges in Existing Graph Library
GraphX
• Not DataFrame based
• Not actively maintained
GraphFrame
• Limited graph
pattern matching
• Semantically weak
graph data model

44. (:Cypher)-[: FOR] -> (Apache:SparkTM)

45. Spark Graph
Given a single Property Graph data model and a
Cypher query, Spark returns a tabular result
[DataFrame]

46. What’s Next?
46
Data Source
APIs
Spark on
Kubernetes
PySpark
Usability
Scala 2.12 Various SQL
Features
GPU-aware
Scheduling
Adaptive
Execution
Hadoop 3.x
Spark Graph
mlflow

47. Data Source API V2
• Unified API for batch and streaming
• Flexible API for high performance
implementation
• Flexible API for metadata management

48. JDBC source with data source v1
df.write.format("jdbc")
.option("url", ...)
.option("dbtable", ...)
.option("driver", ...)
.save()
1. Specify the info of remote catalog for each op

49. JDBC source with data source v1
CREATE TABLE tab1(...) USING jdbc OPTIONS("url" ..., "dbtable" ..., ...)
CREATE TABLE tab2(...) USING jdbc OPTIONS("url" ..., "dbtable" ..., ...)
SELECT * FROM tab1 join tab2
INSERT INTO tab1 SELECT ...
2. Register each table before usage

50. JDBC source with data source v2
New: Register the catalog before usage
spark-defaults.conf
spark.sql.catalog.jdbcCatalogName my.jdbc.v2.impl
spark.sql.catalog.jdbcCatalogName.url …
Spark.sql.catalog.jdbcCatalogName.driver …

51. JDBC source with data source v2
• No need to register the tables.
• Access the tables using n-part name.
• DDL/DML support.
CREATE TABLE jdbcCatalogName.db1.t1(...)
ALTER TABLE jdbcCatalogName.db1.t2 CHANGE COLUMN ...
SELECT * FROM jdbcCatalogName.db2.t3
INSERT INTO jdbcCatalogName.db3.t4 SELECT ...

52. What’s Next?
52
Data Source
APIs
Spark on
Kubernetes
PySpark
Usability
Scala 2.12 Various SQL
Features
GPU-aware
Scheduling
Adaptive
Execution
Hadoop 3.x
Spark Graph
mlflow

53. 53

54. Adaptive Query Processing
Intel Blog: https://tinyurl.com/y3rjwcos
Based on statistics of the materialized plan nodes, re-
optimize the execution plan of the remaining queries
• Self tuning the number of reducers
• Adaptive join strategy
• Automatic skew join handling

55. Adaptive Query Processing

56. What’s Next?
56
Data Source
APIs
Spark on
Kubernetes
PySpark
Usability
Scala 2.12 Various SQL
Features
GPU-aware
Scheduling
Adaptive
Execution
Hadoop 3.x
Spark Graph
mlflow

57. Native Spark App in K8S
• Support for using a pod template to
customize the driver and executor pods.
• Dynamic resource allocation and
external shuffle service.
• Better support for local application
dependencies on client machines
• Driver resilience for Spark Streaming
• Better scheduling support.
57
on

58. 58

59. The other targets in Apache Spark 3.0
• Hadoop 3.x support
• Hive execution from
1.2.1 to 2.3.4
• Scala 2.12 GA
• Better ANSI SQL
compliance
• PySpark usability
59
Please follow the announcements
in Spark + AI Summit @ SF

60. What’s Next?
60
Data Source
APIs
Spark on
Kubernetes
PySpark
Usability
Scala 2.12 Various SQL
Features
GPU-aware
Scheduling
Adaptive
Execution
Hadoop 3.x
Spark Graph
mlflow

61. Apache Spark 3.x
61
Catalyst Optimization & Tungsten Execution
SparkSession / DataFrame / DataSet APIs
SQL
Spark ML
Spark
Streaming
Spark
Graph
3rd-party
Libraries
Spark CoreData Source Connectors

62. Apache Spark™
• Use Cases
• Research
• Technical Deep Dives
AI
• Productionizing ML
• Deep Learning
• Cloud Hardware
Fields
• Data Science
• Data Engineering
• Enterprise
5000+ ATTENDEES
Practitioners:
Data Scientists, Data Engineers,
Analysts, Architects
Leaders:
Engineering Management, VPs,
Heads of Analytics & Data, CxOs
TRACKS
databricks.com/sparkaisummit

63. 63
Nike: Enabling Data Scientists to bring their Models to Market
Facebook: Vectorized Query Execution in Apache Spark at Facebook
Tencent: Large-scale Malicious Domain Detection with Spark AI
IBM: In-memory storage Evolution in Apache Spark
Capital One: Apache Spark and Sights at Speed: Streaming, Feature
management and Execution
Apple: Making Nested Columns as First Citizen in Apache Spark SQL
EBay: Managing Apache Spark workload and automatic optimizing.
Google: Validating Spark ML Jobs
HP: Apache Spark for Cyber Security in big company
Microsoft: Apache Spark Serving: Unifying Batch, Streaming and
RESTful Serving
ABSA Group: A Mainframe Data Source for Spark SQL and Streaming
Facebook: an efficient Facebook-scale shuffle service
IBM: Make your PySpark Data Fly with Arrow!
Facebook : Distributed Scheduling Framework for Apache Spark
Zynga: Automating Predictive Modeling at Zynga with PySpark
World Bank: Using Crowdsourced Images to Create Image
Recognition Models and NLP to Augment Global Trade indicator
JD.com: Optimizing Performance and Computing Resource.
Microsoft: Azure Databricks with R: Deep Dive
Airbnb: Apache Spark at Airbnb
Netflix: Migrating to Apache Spark at Netflix
Microsoft: Infrastructure for Deep Learning in Apache
Spark
Intel: Game playing using AI on Apache Spark
Facebook: Scaling Apache Spark @ Facebook
Lyft: Scaling Apache Spark on K8S at Lyft
Uber: Using Spark Mllib Models in a Production
Training and Serving Platform
Apple: Bridging the gap between Datasets and
DataFrames
Salesforce: The Rule of 10,000 Spark Jobs
Target: Lessons in Linear Algebra at Scale with
Apache Spark
Nationwide: Deploying Enterprise Scale Deep
Learning in Actuarial Modeling at Nationwide
Workday: Lesson Learned Using Apache Spark

64. Thank you
Xiao Li (lixiao@databricks.com)
Wenchen Fan (wenchen@databricks.com)
64